{"id":"2cd1513a-9e0e-45b7-9cd2-a51518d7aa4f","arxiv_id":"1908.08297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"EGNet couples salient edge features with salient object features in a single network and reports top scores on six salient object detection benchmarks.","lead":"This paper presents EGNet, a neural network for salient object detection that explicitly extracts and fuses edge information to sharpen object boundaries. The method reports state-of-the-art accuracy on six saliency benchmarks without pre- or post-processing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablations in Sec. 4.3 add capacity and deep supervision along with edge guidance, so Table 3 does not establish that the reported gains come from the edge cue rather than from extra parameters and losses.","rationale":"The reader's CONDITIONAL verdict is appropriate. I agree with the weakest assumption and focus on the capacity confound because the strongest claim includes a causal claim about edge guidance: the method is said to benefit from complementary salient edge information. The addition of the edge branch and one-to-one sub-side paths changes architecture capacity, the number of supervised side outputs, and the loss-weight balance simultaneously. Without a parameter-matched control, Table 3 cannot attribute the observed gains to the edge cue. The paper also omits how the salient edge ground truth in Eq. (4) is derived, which is a genuine reproducibility gap, but it is secondary: even a well-specified edge target would not resolve the missing capacity control. The availability of source code is a positive independent support, and the Table 2 SOTA comparison is broadly consistent with the field, so no verdict change is warranted. The authors should either supply a capacity-matched ablation or soften the causal interpretation.","tokens_in":14454,"tokens_out":13361,"duration_ms":147008,"concrete_test":"Train a capacity-matched control on DUTS-TR: take baseline B and add exactly the same T/T' convolution layers and side-output branches as in B+edge TDLP+MRF OTO, but omit the saliency-edge loss and edge-guidance fusion. For example, run the added S(2) branch through the same T layers with a duplicate saliency-object supervision instead of Eq. (4), and add the one-to-one sub-side paths with saliency supervisions only, tuning loss weights with the same protocol. Evaluate on SOD and DUTS-TE under the same metrics. If this control matches rows 3 and 6 of Table 3, the edge-guidance attribution is not supported; if it falls clearly short, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim includes a causal statement: edge guidance produces sharper boundaries and better localization. The ablations in Sec. 4.3 do not isolate that cause. Baseline B in Table 3 is a U-Net with side paths; B+edge TDLP and the full OTO model add a new S(2) edge branch, a top-down propagation path, an extra edge loss in Eq. (4), and four one-to-one sub-side paths (each with T' and D' layers, Sec. 3.3) plus their losses. Row 4, B+edge NLDF, is not a parameter-matched control; it only changes the loss of the baseline. Thus the improved F-measure, MAE, and S-measure on SOD and DUTS-TE in rows 2, 3, and 6 could be due to additional capacity, extra deep-supervision signals, or loss-weight balance rather than to the complementary edge features. The SOTA numbers in Table 2 may stand, but the causal story in the abstract and Sec. 5 is under-supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EGNet, a fully convolutional network for salient object detection that explicitly models salient edge information alongside salient object information. The architecture has three components: a progressive salient object feature extraction module (PSFEM), a non-local salient edge feature extraction module (NLSEM) that combines local edge cues from Conv2-2 with top-down location propagation, and a one-to-one guidance module (O2OGM) that fuses the edge features with multi-resolution object features. The network is trained with deep supervision on both edge and object side outputs. Experiments on six benchmarks (ECSSD, PASCAL-S, DUT-OMRON, HKU-IS, SOD, DUTS-TE) report state-of-the-art results under max F-measure, MAE, and S-measure, with and without a ResNet backbone, and without pre- or post-processing. The paper also includes ablations on SOD and DUTS-TE in Table 3.","tokens_in":14724,"tokens_out":4014,"duration_ms":37189,"significance":"If the central claim holds, EGNet demonstrates that explicitly modeling and fusing salient edge features improves both boundary quality and localization in salient object detection, with consistent gains across six datasets and three metrics. The paper's strengths are the breadth of evaluation (15 competing methods, six datasets, three metrics, two backbones), the release of source code, and the use of standard external evaluation protocols. The conceptual contribution—treating salient edge detection and salient object detection as complementary tasks in one network—is timely and has influenced subsequent work. However, the empirical attribution of the gains to edge guidance is currently under-supported by the ablation analysis, as detailed in the major comments.","major_comments":[{"comment":"The ablations do not control for parameter count or for the number of auxiliary losses, so the reported gains cannot be cleanly attributed to the edge-guidance mechanism. Row 6 (B + edge TDLP + MRF OTO) adds, relative to row 1 (B): the S(2) edge branch with its T layers, the top-down propagation convolution in Eq. (2), the extra edge loss in Eq. (4), four sub-side paths each with new T' and D' layers, four additional side losses in Eq. (8), and the fused-map loss in Eq. (9). Any of these additions—especially extra capacity and deep supervision—could explain part or all of the improvements in F-measure, MAE, and S-measure. The comparison to row 4 (B + edge NLDF) is also not parameter-matched, since row 4 only adds an IOU loss to the baseline without adding any of the new branches. To support the causal story in the abstract and Section 5, the authors should add a parameter-matched control (e.g., a baseline with the same number of added convolutional layers and auxiliary losses but without edge supervision) and an ablation that adds the OTO sub-side paths without the edge features.","section":"Section 4.3, Table 3"},{"comment":"The edge supervision in Eq. (4) requires a salient-edge ground-truth label set Z+ and Z−, but the paper never specifies how this edge ground truth is generated from the saliency masks. This is a load-bearing detail for reproducibility and for interpreting the edge-guidance results: different edge extraction procedures (e.g., morphological boundary extraction, Sobel-like filtering, or manual annotation) could substantially change both the training signal and the reported edge-quality numbers in Table 4. Please specify the exact procedure, including any morphological thinning/thickening operations and parameter settings, and if possible report sensitivity to this choice.","section":"Equation (4) and Section 3.2.2"},{"comment":"The claim that the improvement from row 1 to row 3 is obtained 'without additional time and space consumption' is not supported by the architecture as described. Row 3 (B + edge TDLP) introduces at least the Trans convolution and the upsampling operation in Eq. (2) and the S(2) edge branch with its T layers, all of which add parameters and computation relative to row 1. Unless the authors intend a different baseline comparison (for example, against a U-Net of matched total capacity), this sentence should be corrected or substantiated with parameter counts and FLOPs for the relevant configurations.","section":"Section 4.3.2"}],"minor_comments":[{"comment":"In the RFCN row, the S-measure value '0852' on PASCAL-S appears to be missing the decimal point and should likely read '0.852'.","section":"Table 2"},{"comment":"The phrase 'fixed sober operator' should be 'fixed Sobel operator' when describing the NLDF loss.","section":"Section 2"},{"comment":"The caption uses 'edge TDLF' while the text and table rows use 'edge TDLP'; these should be made consistent.","section":"Table 3 caption"},{"comment":"The summation notation in Eq. (10) is awkward as printed ('i=6∑ i=3'); it should be written as a standard sum over i from 3 to 6.","section":"Equation (10)"},{"comment":"All experimental results are reported from a single training run with no error bars or significance tests. Given that the central claim rests on small numerical differences (e.g., Table 3 row-to-row gaps of 0.5–1.5%), a note on run-to-run variance or at least multiple-seed evaluation would strengthen confidence in the results.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nEGNet is worth your time if you work on salient object detection. The reported numbers are strong: it beats PiCANet by about 1.9 F-measure on average across six datasets, without pre/post-processing, and it wins on MAE and S-measure as well. The novel bit is real: a dedicated salient-edge branch built from Conv2-2 plus top-down location propagation, fused one-to-one into each object side path. I do not know of prior work that does exactly that; earlier methods mostly used edge-like losses or separate contour models. The paper also ships code, and the evaluation uses standard benchmarks and published metrics, so the headline comparison is reproducible in principle.\n\nThe soft spot is the causal claim, not the leaderboard. The Section 4.3 ablation adds capacity and deep supervision at the same time as edge guidance: the full model includes the new edge branch, top-down propagation, extra sub-side paths with their own T' and D' layers, and several additional losses, while the baseline has none of those. So Table 3 does not prove that the gains come from the complementary edge features rather than from more parameters and more supervision. B+edge NLDF is not a parameter-matched control either. I would want a matched-complexity ablation, or at least an explicit admission that edge guidance and capacity are confounded. The unspecified edge ground-truth generation for Eq. (4) is a smaller but real reproducibility gap, and the paper reports no error bars or significance tests. Those are fixable in a revision and do not make me doubt the ranking, since the gains are consistent across six datasets and the full model is trained on DUTS without fitting test sets.\n\nDo not overcorrect: the central empirical claim—EGNet is state of the art on these benchmarks at the time—holds up as reported. It is the attribution to edge guidance that is under-supported. This is a legitimate incremental contribution, not a breakthrough, and it reads like an honest ICCV-stage submission. The citations cover the relevant edge-aware and multi-scale SOD methods; the S-measure comes from the same group but is a published external metric, so I do not see a circularity problem.\n\nWho gets value: SOD researchers comparing architectures, and anyone doing multi-task feature fusion. I would cite it as prior work on edge-guided saliency. It deserves a serious referee, and conditional acceptance with requests for a matched-complexity ablation, edge-GT details, and ideally error bars. Not a desk reject.","headline":"Solid incremental SOD architecture with genuine edge-guidance novelty and strong benchmark numbers, but the ablation story does not isolate edge guidance from added capacity.","tokens_in":15208,"tokens_out":2205,"would_cite":true,"duration_ms":24223,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Explicitly modeling edges alongside objects yields the best reported saliency results on six benchmarks.","keywords":["salient object detection","edge guidance","complementary features","fully convolutional network","deep supervision","top-down location propagation","one-to-one guidance module","boundary refinement"],"falsifier":"Train a matched-capacity baseline that adds the same convolutional side branches and the same number of deep-supervision losses but without edge features or the one-to-one guidance; if it ties EGNet on DUTS-TE and SOD under F-measure and MAE, then the edge-guidance claim fails. Alternatively, re-run the ablation with the edge branch fed random noise features of the same shape; if performance does not drop, the guidance is not carrying information.","tokens_in":14295,"feed_emoji":"🎯","tokens_out":4659,"duration_ms":44903,"temperature":0.7,"pith_summary":"This paper argues that salient-object detection is held back by ignoring a complementary signal: the salient edge. It introduces EGNet, a single end-to-end network that explicitly predicts salient edges and salient objects at the same time, then lets the edge features guide the object features at several resolutions. The claim is that this complementary modeling improves both the sharpness of object boundaries and the accuracy of object localization. On six standard datasets, the reported F-measure, mean absolute error, and S-measure are the best among the compared methods, without any pre- or post-processing. A sympathetic reader would take the central discovery to be that edge information is not just a boundary polish but a lane marker that helps find the object.","feed_headline":"Edge guidance sharpens salient object boundaries","feed_subtitle":"A single network joins edge and object prediction to lift localization and boundary quality on six benchmarks.","key_machinery":"The load-bearing mechanism is a pair of modules. The non-local salient edge features extraction module takes the low-level layer that best preserves edge detail and adds to it a top-down propagated signal from the deepest, most location-aware layer, then supervises the result with a salient-edge loss; this yields edge features that respond only to edges belonging to salient objects, not background clutter. The one-to-one guidance module adds those same edge features into every resolution of the object-feature pyramid before further convolution and deep supervision, so the edge cues are not diluted by progressive fusion. The whole system is trained end-to-end with cross-entropy losses on every side output plus the fused output.","core_discovery":"The central claim is that salient edge information and salient object information are complementary, and that modeling both in one network, with the edge features propagated to guide the object features, yields sharper saliency maps and better localization simultaneously. The paper states this as three steps: progressive fusion of multi-scale object features from a backbone; extraction of salient edge features by combining local edge cues from a low-level layer with top-down global location information; and a one-to-one guidance module that fuses the same edge features into each resolution of object features. With joint supervision on both tasks, the final fused prediction is reported to outperform 15 previous methods on ECSSD, PASCAL-S, DUT-OMRON, HKU-IS, SOD, and DUTS-TE, under max F-measure, MAE, and S-measure, without pre-processing or post-processing.","pith_inferences":["The unspecified salient-edge ground truth, presumably derived as boundaries of the saliency mask, makes the method a form of self-supervision; a testable extension is to verify whether richer human-annotated contours change the gain.","The same complementary-guidance design could transfer to other paired prediction problems, such as semantic segmentation with contour heads or depth estimation with edge-aware refinement; the paper does not test this.","If the benefit persists when the edge branch is replaced with a fixed edge detector such as a morphological gradient, then the gain would come from the location-propagation path rather than learned edge semantics; the paper does not run this control.","The comparison against a loss-level edge penalty suggests that feature-level guidance, not loss shaping, drives the improvement; an untested corollary is that feature-level guidance would also help when applied to that penalty-based baseline."],"forward_implications":["Saliency maps from the fused output have sharper boundaries and more accurate localization than the baseline that fuses only object features, without any extra inference-time processing.","On six standard benchmarks, the reported numbers are the best among the compared methods, so a method that wants to claim state of the art now has EGNet as the reference point.","The edge branch provides an additional supervision signal during training, so the object branch is trained with richer information than it would get from saliency masks alone.","The one-to-one fusion beats both fusing edge features at the top only and progressive upward fusion, so how edge cues are injected matters as much as whether they are injected."],"supporting_citations":[{"why":"NLDF, the IOU edge-loss baseline the paper argues against; provides the edge-penalty comparison and the salient-edge outputs used in Table 4.","marker":"[35]"},{"why":"DSS, the deep-supervision and short-connection architecture whose setup the backbone follows and a main deeply-supervised competitor.","marker":"[17, 18]"},{"why":"U-Net, the encoder-decoder structure on which the progressive salient-object feature extraction is built.","marker":"[40]"},{"why":"PiCANet, the strongest prior method; its reported numbers are the direct comparison for the state-of-the-art claim.","marker":"[33]"},{"why":"The S-measure metric definition used in all evaluations.","marker":"[10]"},{"why":"DUTS, the dataset used for training and for the large-scale test set DUTS-TE.","marker":"[46]"}],"fun_headline_variants":["Edge guidance sharpens salient object detection","EGNet nets sharper boundaries via edge fusion","Complementary edge cues boost saliency maps","Salient edges guide object detection in one pass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the measured gains come from the edge-guidance mechanism itself; the paper's ablations do not hold parameter count, number of side outputs, or loss-weight budget fixed, and the salient-edge ground truth used for supervision is never specified.","fun_headline_variants_meta":{"raw":{"variants":["Edge guidance sharpens salient object detection","EGNet nets sharper boundaries via edge fusion","Complementary edge cues boost saliency maps","Salient edges guide object detection in one pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000615,"raw_usage":{"total_tokens":2838,"prompt_tokens":908,"completion_tokens":1930,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":1875}},"tokens_in":524,"tokens_out":1930,"duration_ms":15045,"temperature":1.0,"reasoning_tokens":1875,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:42:35.503927+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a matched-capacity baseline that adds the same convolutional side branches and the same number of deep-supervision losses but without edge features or the one-to-one guidance; if it ties EGNet on DUTS-TE and SOD under F-measure and MAE, then the edge-guidance claim fails. Alternatively, re-run the ablation with the edge branch fed random noise features of the same shape; if performance does not drop, the guidance is not carrying information.","supporting_citations":[{"cited_title":"Non- local deep features for salient object detection","cited_arxiv_id":null,"evidence_quote":"NLDF, the IOU edge-loss baseline the paper argues against; provides the edge-penalty comparison and the salient-edge outputs used in Table 4."},{"cited_title":"Picanet: Learning pixel-wise contextual attention for saliency detec- tion","cited_arxiv_id":null,"evidence_quote":"PiCANet, the strongest prior method; its reported numbers are the direct comparison for the state-of-the-art claim."},{"cited_title":"Structure-measure: A new way to evaluate foreground maps","cited_arxiv_id":null,"evidence_quote":"The S-measure metric definition used in all evaluations."},{"cited_title":"Learning to de- tect salient objects with image-level supervision","cited_arxiv_id":null,"evidence_quote":"DUTS, the dataset used for training and for the large-scale test set DUTS-TE."}],"review_version":1}